Building extractionhome.iitk.ac.in/~blohani/LiDARSchool2008/Downloads/...International School on...

International School on LiDAR technology, IIT Kanpur, 31 March -4 April 2008.

Building extractionBuilding extraction

Dr. Bharat Lohani D t t f Ci il E i iDepartment of Civil EngineeringIIT Kanpur

Mr. Rajneesh Singh, M.Tech.

Why building extraction?

Update GIS databaseUpdate GIS database3D city modelMap makingMap makingRevenue collectionChange detectionChange detectionDisaster management

International School on LiDAR Technology, IIT Kanpur, 31 March – 4 April 2008

Where are buildings in the point cloud?

Manual interpretation !

How a computer knows that a group knows that a group of point cloud is actually for a building?


Desired outcomePlanPlanHeight3D vector model3D vector model3D vector model with texture


Why difficult ?

In LiDAR data no other but geometric In LiDAR data no other but geometric information in the form of sparse pointsGeometry alone can separate from Geometry alone can separate from natural features, e.g., trees but not from the artificial features, e.g., tanks or double decker busThe buildings may be simple and very complexcomplexThere are errors in data


Basic concepts…

Height difference wrt local neighborhoodHeight difference wrt local neighborhoodTypical shapes different from natural objectsjMade of planar surfacesDomain knowledge


Level of detail


Possible approach: Using only laser data

Based on detection of planes Based on detection of planes Edges using plane intersection or height difference in DSMdifference in DSMLimitations of using only LiDAR data

Confusion in trees and buildingsConfusion in grass lands and roadsAs the only criterion of classification is the height or geometry (plane)g y (p )


Possible approach: Using laser and Possible approach: Using laser and other data

Aerial imagesAerial imagesSatellite imagesGround plans Ground plans Advantages of both data sets

To eliminate the limitations of laser To eliminate the limitations of laser alone


DSM, BEM and nDSM


Approach: Using thresholded nDSM

Generate BEM (DTM)Generate BEM (DTM)Find normalized DSM = DSM – DTMThis normalized DSM shows the variation of only above ground objects as terrain slope has been above ground objects, as terrain slope has been eliminated.Threshold normalized DSM at different heights. The irregular shapes corresponds to trees while The irregular shapes corresponds to trees while regular shapes corresponds to buildings. A process of eliminating trees and retaining buildings leads to finding outlines of buildings leads to finding outlines of buildings.

Morphological operatorsTexture analysis

Find planes and model and group


Find planes and model and group

Approach: Concept of horizontal profiles

The size and CG location of building will not change much in comparison to other items as we slice upwards


items, as we slice upwards.

Building identification by variation in g ysize and location of CG of blobs

+ Domain knowledge, other dataInternational School on LiDAR Technology, IIT Kanpur, 31 March – 4 April 2008

+ Domain knowledge, other data

Approach: ClassificationUsing classification of LiDAR Using classification of LiDAR DSM and spectral dataGenerate nDSMDSM f t th nDSM forms yet another

band along with the spectral data for classificationT di i i t t f To discriminate tree from grass, concrete road from concrete roof…as height is also therealso thereTo discriminate building from tree as thematic information is there


information is there

Spectral data: Digital Spectral data: Digital imageConvert spectral data (d l )(digital image) into orthophotograph for registration with the LiDAR data


Classification procedurenDSM is considered nDSM is considered as one of the bandsThe CIR band used f d lfrom digital imageISODATA classification used for generating clusters


Approach: Using planar surfaces

B ildi d f l fBuildings are made of planar surfacesLocate planes within point cloudIntersect planes to determine edgesUse local height variation to distinguish from non-building planes like roads, grounds etc.


Hough transform 2Dy

y mx c= +y

c

x( , )m c Line in Cartesian space

Line in parameter space


m

Hough transform 2Dyy

c

1 1( , )x y

xPoint in Cartesian space

1 1c x m y= − +Line in parameter space


mLine in parameter space

Hough transform 2Dyy

2 2( , )x y

c

1 1( , )x y

x

c x m y= − +Point in Cartesian space

1 1c x m y= +

2 2c x m y= − +


m Line in parameter space

Hough transform 2Dy ( )x yy

2 2( , )x y3 3( , )x y

c

1 1( , )x y

x1 1c x m y= − + Point in Cartesian space

2 2c x m y= − +

3 3c x m y= − +


mLine in parameter space

Hough transform 2D

yUsing the above philosophy the parameter space for a set of points will look like:

c

x

So can we use the density in So can we use the density in parameter space to see that for what values of ‘m’ and ‘c’ there is a line.


m

Hough transform 2D

yDetermine parameters for all possible neighboring pairs.

c0 5 2 2

x

Populate discretized parameter

0 5 2 2

1 27 4 3

Populate discretized parameter space with the counts of similar ‘m’ and ‘c’.

1 3 1 3

5 2 1 0


m

Hough transform 2D

yThreshold the Hough room.

c

x

Locate the point-pairs in Cartesian

27

Locate the point pairs in Cartesian space which have populated the high frequency cell of parameter space.


m

Hough transform 3D

Building planes are in 3D space and need to be detected from 3D LiDAR point cloud.

Need to apply Hough transform in 3D


Hough transform 3D

LiDAR data for roof top


Like the line segments of 2D here triangle is the basic element.

z

c

bx

yc

z ax by c= + +x

Planes in Cartesian space

a b

Plane in parameter space


Plane in parameter space

zz

•Create 2D TIN on LiDAR data (x,y)

•Consider height of each vertexz ax by c= + +

x

y

•Determine the parameters (a, b, c)

of all planes.

•Populate the 3D Hough room with

z ax by c= + +c

•Populate the 3D Hough room with

the count, i.e., how many planes

have similar parameters

•Threshold 3D Hough room

a b


b

•The triangles for which parameters

are nearly same and with high

frequency in Hough room

z

cz

x

y

a b


Observations

The result is identification of all LiDAR The result is identification of all LiDAR points that form planar surfaces. One Hough room may represent several One Hough room may represent several roof tops…

How is this affected by:The accuracy of dataThe accuracy of dataData density


Point classification as planesHough room method classifies LiDAR points as planar Hough room method classifies LiDAR points as planar surfaces.


Building modelling

R i d i t t t d l ( i f f Required is to generate a vector model (wireframe of building)


Separation of roofs from one major p jplane (one hough room cell)

•Rasterize the points

•Use image processing algorithms 74for blob identification

•Give different IDs to blobs

•Map the blobs back in LiDAR point 24

4

•Map the blobs back in LiDAR point

cloud to give different IDs to

different point groups

2

•Remove smaller <threshold planes

•Merge closer groups


Plane edge identification•Rasterize point groups Rasterize point groups

•Morphological closing

•Edge identification


Line fitting to edge data points


Result


Effect of data type on building extraction

lue

Flying height Scan angle

index

va

Acc

ura

cy

A


Data density

Other planer approaches…

S t ti i f lSegmentation using surface normalsRANSAC (Random ConcensusS li ) th dSampling) methodOthers…Terrascan uses holes in ground class to locate possible building points


ReferencesBreener, C., 2005, Building reconstruction from images and laser scanning , International , , , g g g ,Journal of Applied Earth Observation and Geoinformation, Volume 6, Issues 3-4, March 2005, Pages 187-198Haala, N. and Brenner, C., 1999. Extraction of buildings and trees in urban environments. ISPRS Journal of Photogrammetry and Remote Sensing, 54(2-3): 130-137.Lohani, B. and Singh, R., 2008, Effect of data density, scan angle, and flying height on the o a , a d S g , , 008, ect o data de s ty, sca a g e, a d y g e g t o t eaccuracy of building extraction using LiDAR data, Geo Carto International, Vol. 23, No. 2, April 2008, 81–94Maas, H.-G. and Vosselman, G. (1999). Two algorithms for extracting building models from raw laser altimetry data. LISPS Journal of Photogrammetry and Remote Sensing, 54:153163.Sohn, G. and Dowman, I. (2003). Building extraction using lidar dems and ikonos images. Sohn, G. and Dowman, I. (2003). Building extraction using lidar dems and ikonos images. Proceedings of the ISPRS working group III/3 workshop ‘3-D reconstruction from airborne laserscanner and InSAR data’ Dresden, Germany.Vosselman, G. (2002). Fusion of laser scanning data, maps and aerial photographs for building reconstruction. IEEE International Geoscience and Remote Sensing Symposium and the 24th Canadian Symposium on Remote Sensing, IGARSS’02, Toronto, Canada,.y p g, , , ,Zhan, Q., Molenaar, M., and Tempfli, K. (2002b). Hierarchial image object based structural analysis toward urban land use classification using high-resolution imagery and airborne lidardata. IProceedings of the 3rd international symposium on remote sensing of urban area’s 2002, pages 251–258.


Building extractionhome.iitk.ac.in/~blohani/LiDARSchool2008/Downloads/...International School on...

Documents

Transcript of Building extractionhome.iitk.ac.in/~blohani/LiDARSchool2008/Downloads/...International School on...